New method lets DeepSeek and other models answer ‘sensitive’ questions

Bias and some cases, open censorship, large language models (LLMS) are difficult to eliminate. Such a model, Depth From China, Alarm politicians and some business leaders about the potential threat to national security.

A committee recently elected in the US Congress left a report Deepseek is called “a deep hazard for the safety of our nation” and detailed policy recommendations.

There are ways related to learning to strengthen human feedback (RLHF) and subtle regulation, enterprise risk management Ctgt requires an alternative approach. CTGT has prepared a method of biased and censored, wrapped in some language models that say that it eliminates 100% censorship 100% censorship.

One paperCyril Gorlla and CTGT’s Trevor Tuttle, the frames “found the internal features responsible for direct censorship,” he said.

“This approach also allows you to delicate model behavior that ensures the delivery of the model, not effective, but without violating the general capabilities and actual accuracy of the model,” he said.

Method is clearly designed with DeepSEek-R1 Distilla-70B-70B, the same process can be used in other models.

“We have tried CTGT with other open-weight models such as LLAM and found it as effective,” Gorlla told Venturebeat in an email. “Our technology applies to all the deep learning models, which applies to basic deep learning models.

How does it work

Researchers said they identify the properties of their methods to be related to unwanted behaviors.

“The key ideas are in a large language model, ‘censorship trigger’ or ‘censorship triggers’ or’ venerable senses that are in the concepts of toxic feelings (in hidden in neurons or directions).

CTGT said there were three main steps:

1. Feature authentication
2. Feature insulation and characteristic
3. Dynamic feature modification.

The researchers are a number of tips that can take one of the poisonous feelings. For example, the Tiananmen Square can ask for more information on or transitions to the transition.

After these are determined, researchers can beolate this feature and manage the part of unwanted behavior. Behavior can reply to more caution or refuse to fully respond. By understanding the behavior of feature control, researchers and feature, the model will be able to integrate a mechanism in the inferences pipeline. “

The model gives you more instructions to answer

CTGT, using 100 sensitive surveys, showed that the DeepSEEK-R1-R1-R1-R1-R1-R1-R1-R1-R1-R1-R1-R1-R1-R1-R1-R1-R1-R1-R1-R1-R1-R1-70B has responded only to 32% of the controversial instructions. However, the modified version responded to 96% of the instructions. The remaining 4%, the CTGT explained was an extremely open content.

The company said that the method allows users to work users how many biased and security features will be operated, but if the model will not turn into a “careless generator”, especially if unnecessary censorship.

His method also does not sacrifice the accuracy or performance of the model.

“This can change between changes or adaptations, between changes or adaptations, between changes or adaptations, which are not optimizing the model weights or new sample response, or for different contexts that can change between different behaviors.

Model safety and security

DeepSEEK reported that the Congress report expands export control, to take a speedy measure to ensure export control and increase the risks from Chinese artificial intelligence models. ”

After the US government questioned the potential threat for DeepSEEK, researchers and EU companies searched for the ways and other models, “safe” ways.

“Safe” or biased or biased or censorship can sometimes be difficult to judge, but can develop methods that allow users to think about how examples will be for them.

Gorlla, enterprises’ said it should be able to Trust their models Matches with their policies, so the methods that help develop it will be critical for business.

“CTGT is specifically important for millions of dollar fashion models for each use for each use, the AI ​​in high-risk applications such as security, finance and health, where the damage caused by AI malfunction is violent.”